Remote Sensing Image Detection Method Combining Dynamic Convolution and Attention Mechanism
Small object detection in remote sensing images is challenging. Traditional CNN downsampling often leads to the loss of small object details and missed detections. This paper proposes an improved YOLOv8 algorithm, incorporating adaptive feature extraction and multi-scale fusion modules to enhance fe...
Saved in:
| Main Authors: | , , |
|---|---|
| Format: | Article |
| Language: | English |
| Published: |
IEEE
2025-01-01
|
| Series: | IEEE Access |
| Subjects: | |
| Online Access: | https://ieeexplore.ieee.org/document/10924169/ |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | Small object detection in remote sensing images is challenging. Traditional CNN downsampling often leads to the loss of small object details and missed detections. This paper proposes an improved YOLOv8 algorithm, incorporating adaptive feature extraction and multi-scale fusion modules to enhance feature representation, effectively capturing the details of small objects. We introduced the AFGCAttention mechanism to strengthen the network’s focus on key regions while suppressing irrelevant background information, improving the model’s ability to recognize small objects. To address the resolution loss issue in small object detection, we adopted the CARAFE (Content-Aware ReAssembly of FEatures) upsampling operator. By performing content-aware reassembly of feature maps, CARAFE avoids the blurriness and information loss commonly associated with traditional upsampling methods, demonstrating significant advantages in reconstructing the details of small objects, resulting in clearer and more accurate boundaries. Additionally, to improve the accuracy of bounding box regression, we integrated the GIoU loss function to optimize the geometric matching between ground truth and predicted boxes, addressing the problem of inaccurate bounding box localization for small objects and enhancing localization precision. Experimental results demonstrate that the proposed algorithm significantly improves the precision of small object detection and maintains robustness in complex backgrounds. The improved model achieved an mAP of 83.0%, with accuracy improvements of 85.0%, 2.0%, and 5.0% compared to the baseline. Compared with existing detection methods, this approach shows outstanding performance in detection accuracy, localization precision, and computational efficiency, particularly excelling in small object detection. |
|---|---|
| ISSN: | 2169-3536 |